Anthropic Launches an Official Plugin Directory for Claude Code

🧠 LAUNCH

Zyphra drops ZONOS2 — a next-generation real-time TTS model with high-fidelity voice cloning, now open-source. The open voice synthesis space keeps accelerating; if you're building anything voice-powered, this is worth benchmarking against your current provider. (670 likes | 101 RTs) Read more →

Rio de Janeiro claims a 397B "homegrown" LLM — a municipal government releasing a model this size would be unprecedented, but the story unraveled fast. Community investigation (see RESEARCH) reveals it's likely a merge of existing open-weight models relabeled as original work. The announcement got 3.7K likes before the debunking caught up. (3,698 likes | 269 RTs) Read more →

🔧 TOOL

Anthropic Launches an Official Plugin Directory for Claude Code

claude-plugins-official hit 30K GitHub stars almost immediately — Anthropic is building an ecosystem layer, shifting Claude Code from standalone agent to extensible platform. Think VS Code's extension marketplace, but for AI coding agents. This is the clearest signal yet that the moat isn't the model, it's the developer experience wrapped around it. If you use Claude Code daily, browse the directory now — the early plugins cover MCP servers, workflow hooks, and domain-specific tooling. (30,126 likes | 3,257 RTs) Read more →

Oh-My-Codex Turns Codex Into a Multi-Agent System

Oh-My-Codex adds hooks, agent teams, HUDs, and orchestration to OpenAI's Codex — the "oh-my-zsh for AI coding agents" pattern has arrived. It hit 30K GitHub stars in days, which tells you how badly developers want extensibility in their coding agents. The hooks system alone — letting you trigger actions on agent events — fills a gap that vanilla Codex doesn't touch. (30,898 likes | 2,428 RTs) Read more →

Xiaomi's MiMo Code quietly dropped a free Claude Code alternative with persistent memory across sessions, subagents, and zero-config setup — and hit 8K GitHub stars in four days. The open-source AI coding agent space is getting genuinely competitive. If you want a backup agent that doesn't depend on any single provider's API access, this is worth a look. (260 likes | 37 RTs) Read more →

📝 TECHNIQUE

Google publishes a 421-page AI agent engineering playbook — covering prompt chaining, memory architectures, MCP integration, multi-agent systems, and guardrails, all with code examples. This is effectively a free curriculum for anyone building agents. Download it now; it's the most comprehensive single resource on agent engineering published this year. (249 likes | 99 RTs) Read more →

The June 2026 guide to running local LLMs on consumer GPUs — a practical llama.cpp walkthrough covering Gemma 4-12B, Qwen, and more with specific VRAM requirements and real performance numbers. Especially timely as developers rethink their API dependencies post-Fable ban. Match your GPU to the right model before you spend a weekend configuring something that won't fit. (628 likes | 45 RTs) Read more →

Codex Mobile power user guide — your phone as a coding control plane. Covers branch selection, worktrees, /side threads, and /plan mode from mobile, all controlling agents running on your desktop. If you're already using Codex, this turns dead time (commutes, waiting rooms) into productive review cycles. (192 likes | 14 RTs) Read more →

🔬 RESEARCH

New Benchmark Pits 7 Frontier Models on Autoresearch Tasks

Finally, an independent benchmark that tests what matters: agentic research capability across ML engineering, prompt engineering, and open-ended research tasks. Seven frontier models compared head-to-head on tasks that require multi-step reasoning, tool use, and synthesis — not just pattern matching. Independent agentic benchmarks are rare, and this one fills a real evaluation gap. Check where your model of choice actually ranks before you bet your pipeline on it. (368 likes | 28 RTs) Read more →

Community investigation exposes Rio's "homegrown" LLM as a model merge — a detailed GitHub issue walks through the evidence that Rio de Janeiro's 397B parameter model is a merge of existing open-weight models, not original training. The weight analysis is damning. A cautionary tale about AI announcements from non-traditional players — and a reminder that as open models proliferate, provenance verification matters as much as benchmark scores. (260 likes | 139 RTs) Read more →

💡 INSIGHT

Simon Willison: Why AI Hasn't Replaced Software Engineers, and Won't

The developer who spent a week converting to Fable's capabilities — building tools, shipping features, going from skeptic to power user — then lost access overnight when the government pulled the plug. His essay isn't the usual "AI is just a tool" handwave. It's a specific, experience-driven argument about why engineering judgment, context management, and system design are the parts AI can't replace. The timing gives it teeth: you don't get to dismiss "AI won't take your job" when the guy saying it just lost his favorite model and is more productive, not less, for understanding why. Share this with your non-technical stakeholders. Read more →

Fable 5 suspension crosses 42K likes as the fallout continues. Anthropic's official suspension tweet keeps climbing — the engagement alone tells you how many developers were building on Fable daily. If you haven't checked your own Fable 5 access status and dependency exposure, do it now. (42,873 likes | 7,068 RTs) Read more →

Simon Willison on the Fable jailbreak: "I'm just glad nobody at the US government thought to try that Fable 5 jailbreak against Opus 4.x or GPT 5.x." The wry implication is serious — the vulnerability that triggered the ban likely exists across frontier models. Banning one model for a class-wide issue doesn't make anyone safer; it just makes one company's users collateral damage. (1,577 likes | 56 RTs) Read more →

State attorneys general open investigation into OpenAI — this time over ad policies and health data handling, not model capabilities. Regulatory pressure is now hitting multiple frontier labs simultaneously. The Fable ban was a capabilities story; this is a data practices story. Different vector, same message: the regulatory environment for AI just got real. Read more →

Amazon CEO raised Anthropic concerns before the government crackdown — the investor-triggering-regulation narrative is now mainstream tech press. When your largest investor is flagging concerns to the government about your model's capabilities, the old playbook of "ship fast, lobby later" doesn't work anymore. Audit your AI vendor relationships for single points of failure. Read more →

Interconnects: Welcome to the AGI era of AI governance. One of the most thoughtful AI newsletters argues we've crossed a one-way door — the Fable ban is exhibit A that governance frameworks designed for narrow AI can't handle frontier capabilities. We weren't ready, and there's no going back to the old equilibrium. Read more →

🏗️ BUILD

Mollick's last Fable project — a one-shot FTL travel simulator built with a single prompt, showcasing the model's creative coding capability at the exact moment it became unavailable. A bittersweet demo: the prompt is elegant, the output is impressive, and you can't run it on the model it was designed for anymore. Try it on available models and compare — the gap (or lack thereof) tells you something about where frontier capabilities actually are. (691 likes | 41 RTs) Read more →

🎓 MODEL LITERACY

Model Merging (SLERP / TIES / DARE): Rio de Janeiro's "397B homegrown LLM" turned out to be a merge of existing open-weight models — which makes this the perfect week to understand how model merging actually works. SLERP (Spherical Linear Interpolation) smoothly interpolates between two models' weight matrices, creating a blend that can inherit strengths from both parents. TIES (TrIm, Elect, and merge Signs) resolves parameter conflicts when merging multiple models by identifying which parameters actually changed during fine-tuning and keeping only the important ones. DARE (Drop And REscale) randomly drops delta parameters and rescales the survivors, acting like dropout but at merge time. These techniques are legitimate research tools — but calling a merged model "homegrown" is like calling a remix an original composition. As open models proliferate, provenance verification becomes as important as benchmark scores.

⚡ QUICK LINKS

Did Anthropic actually ask for the Fable ban?: Hacker News–trending analysis questions whether Anthropic welcomed the intervention. (119 likes | 81 RTs) Link
KPMG pulls AI report after its own AI hallucinated the data: The irony writes itself. Link
LeCun: Americans will have walled AI gardens: The Fable ban as proof that API-dependent AI is access-as-privilege. (286 likes | 61 RTs) Link
HuggingFace CEO: Two paths for AI — closed APIs or open models: Clement Delangue frames the post-ban choice. (895 likes | 98 RTs) Link
The Register: AI is code, not magic — stop trying to prompt it smarter: A useful framing for managing stakeholder expectations. (45 likes | 18 RTs) Link
Mollick: Nobody actually knows how to restructure companies around AI agents: Refreshing honesty — practical agents are 6 months old. (294 likes | 23 RTs) Link

🎯 PICK OF THE DAY

Simon Willison's case for why AI won't replace engineers lands harder than any think piece could. The developer who spent a week converting to Fable's capabilities — building tools, shipping production features, publicly documenting his journey from skeptic to power user — then lost access overnight when the US government pulled the plug. And his response wasn't panic or despair. It was the clearest articulation yet of what engineering actually is: the judgment calls, the context that doesn't fit in a prompt, the system design decisions that compound over months. What makes this different from the usual "AI is just a tool" reassurance is that Willison isn't theorizing — he's writing from the specific experience of losing his most capable AI tool and discovering he's more productive, not less, for understanding why it helped. His timing makes the argument undeniable: you can't dismiss "AI won't take your job" when the person saying it just had the rug pulled and is still standing. For every engineer anxious about being replaced, and every manager wondering if they should be, this is the piece to read this weekend. Read more →

Until next time ✌️